Most optimisation moves the schedule. We read what your line already records, recover the physics underneath it, and hand back a process that is different — geometry, sequence, thermal budget, tolerance stack.
Every station on a modern line is already producing more signal than the people running it can consume. Historian tags at one hertz. Vision frames nobody looks at twice. Vibration, ambient, torque curves, lot genealogy.
None of it is missing. It is unexamined, and it is examined one variable at a time when it is examined at all.
A process specification carries hundreds of numbers. A handful of them are load-bearing: they set the tolerance stack, the cycle floor, the thermal ceiling. The rest are inherited, copied from the previous line, or set once by someone who has since left.
We recover which is which from the data rather than from the document, because the document records the intent and the data records the process.
Once the governing constraints are known, the sequence can be searched rather than assumed. Steps that were serial because of a fixture, not because of physics. A cure time set for a material that was requalified four years ago. A gauge station measuring something the previous station already guarantees.
The search space is large, which is the reason to do it numerically instead of in a workshop.
Not a dashboard. A proposed revision to a named parameter, the physics that justifies it, the historical window it was validated against, and the failure mode it is expected to move.
Your process engineer signs it or refuses it. That is the only gate that matters, and we do not sit on the far side of it.
The arm reaches into the magazine, picks a fastener, carries it to the head, drives it home and comes back. Nine moves and three waits, over and over. Every one of those waits is already timed by the controller, and none of them has been looked at since the program was written.
Scroll, and the output gear is replaced by one with fewer teeth. The ratio, the speed and the cycle time beside it are derived from the parts on screen, not typed. That is what a physically justified change looks like: one number, and everything downstream follows.
Clamp, inject, hold, cool, eject. The cool phase is half the cycle and it was set once, for a material that has since been requalified. The dashed bar is the same cycle with the cool phase the data supports. The chart is why.
Five tolerances on one critical dimension. The arithmetic is the worked example from our own article: one contributor carries most of the variance, and the tight ones cost money for a few percent of the outcome. The drawing calls out the two that can be drawn.
A revision is small. Twelve of them are not. The band is three illustrative per-revision rates compounding; where a line lands inside it depends on how many binding parameters it has and how many have been examined. None of these curves is a client figure, and the chart says so.
The magazine is presented where the cycle ends rather than where the bench happened to be free, so the reach is a little over half what it was. The wait at the fastener is what the torque curve needs instead of what was set by habit. Nothing else moved, which is the point: you can see what changed because everything else is where it was.
Historian, PLC tags, MES lots, vision, vibration, maintenance logs. No writes, no line changes, no new hardware in the cabinet. Everything lands on one clock.
Fit the constraints that actually bind this cell, and separate them from the settings that merely persist. The output is a model of the process, not a correlation table.
Hold the recovered constraints fixed and vary everything else: sequence, geometry, dwell, thermal budget, gauge placement. Reject anything the physics forbids before a human ever sees it.
One named change, its justification, the window it was validated against, and its expected failure mode. Reviewed and signed by your engineer, or refused by them.
No. The connection is read-only and stays read-only. Nothing we deploy has write access to a PLC, a historian or an MES, and no proposed change reaches the floor except through your own change-control process.
Not unless you decide it should. The default deployment is on-premise and air-gapped: the model runs beside the historian and nothing egresses. A hosted option exists for sites that prefer it, and it is opt-in.
That is usually the finding worth having. We show the evidence, the window it holds over, and where the specification and the observed process diverge. What to do about it is an engineering decision, and it is yours.
The first constraint model comes from historical data, so it does not wait on new production. Timing depends on how much history is retained and how cleanly it is tagged, which we assess before anyone signs anything.
A walkthrough is a working session, not a demo. Bring the tag list and the specification for a single station and we will show you what can be recovered from it, and what cannot.